Generalizing From Examples
Overview
Users describe by listing instances. "Like X" or "for example Y" means they trust you to find the larger concept.
Two imperatives; which leads depends on context:
| Context |
Lead With |
Why |
| Architecture / design |
Abstract first |
Concept determines what to build. "Add monitoring, for example track response times" → Observability, not latency tracking. |
| Research / clarification |
Enumerate first |
The list IS the deliverable. Concept organizes it. |
Enumeration is always produced. The context determines which imperative drives.
When to Use
Triggers:
- Example markers: "for example", "e.g.", "like", "such as", "including", "things like", "something like"
- Softening qualifiers: "or something", "and so on", "etc.", "and things like that"
- Tool-as-category: "use something like Redis" → caching, not Redis
- One concrete case where the domain obviously contains many parallel cases
Do NOT use when:
- Closure language: "exactly", "specifically", "only", "precisely"
- Closed-set signals: numbered lists, cardinal words ("the three things"), final "and" without "etc."
- Inherently singular scope (one bug, one file, one function)
- You're guessing categories the user wouldn't recognize as related
Core Pattern
The Inference Ladder
- Detect marker — "for example" = illustrative, not exhaustive
- Isolate example — what concept is this an instance of?
- Identify concept — what ONE thing do all potential instances share?
- Infer scope — architecture: all items across system; research: all items in category
- Enumerate cases — always produced, never skipped
- Confirm boundary — "Does this cover what you meant?"
The ladder is recursive. Re-climb on your own output. The most common failure is rename-as-generalize: your "concept" is just the example in disguise. Test: can you name instances the example didn't give? If not, re-climb.
Closure vs. Openness Signals
The user signals either a closed set (complete) or an open set (samples). Look for the signal, not the format.
Closure (exhaustive — do NOT add):
| Signal |
Examples |
| Numbered lists / cardinal words |
"1. auth, 2. rate limit", "the three things are" |
| Definite quantifiers / finality |
"the complete list", "nothing else", "that's it" |
| Closing conjunctions |
"X, Y, and Z" (final "and" with NO "etc.") |
| Closure words |
"specifically", "only", "exactly", "just" |
Openness (illustrative — generalize freely):
| Signal |
Examples |
| Example markers |
"for example", "e.g.", "like", "such as" |
| Softening qualifiers |
"or something", "and so on", "etc." |
| Indefinite / hedging |
"some", "a few", "maybe", "could include" |
| Tool-as-category |
"use Redis" when context is caching strategy |
When both present: primary requester's directive wins. Another stakeholder's dimension is a competing scope definition, not another example — document the conflict, don't resolve by inclusion. Legal requirements (GDPR, SOC2) get flagged separately. Drift test: remove the primary requester's message — would your output still make sense? If yes, you've drifted.
Abstraction Mechanics
The concept is the invariant across all instances — the underlying capability, not a category label.
Abstraction depth:
| Example (shadow) |
Widening (still shadows) |
Concept (the form) |
| "track response times" |
"monitor all endpoints" |
Observability: golden signals across all services |
| "use something like Redis" |
"pick a cache" |
Caching layer strategy: session store, query cache, rate limiter |
| "validate email format" |
"validate other fields" |
Input integrity: format, type, range, sanitization, business rules |
Acid test: present the concept without the example — would the user recognize it?
Confidence: 1 example = confirm before enumerating widely. 2–3 = enumerate as educated inference. 4+ = the abstraction IS the request.
Enumeration Rules
- Name the abstract category
- List candidates with user's examples nested within
- Frame additions as inference: "Based on [category], also consider [cases]."
- If abstraction might overreach: "Do you mean [deeper concept]? If so, I'd cover [cases]."
- Let the user trim; don't self-censor. Never replace the user's examples — build outward.
Structure by category, not by example. Example-anchored output means you've appended, not generalized.
❌ Caching Research → Priority 1: Redis (user mentioned), Priority 2: Memcached
✅ Caching Layer Strategy → Session store: Redis, Memcached; Query cache: Redis (read-replica); Object cache: Hazelcast
Test: can a reader identify which items the user mentioned? If yes, restructure.
Stop and Re-climb
| What You're Doing |
Why It's Wrong |
| Treating "for example X" as "only X" |
The marker signals non-exhaustiveness. X is a starting point. |
| Asking the user to enumerate |
The marker means the user expects YOU to enumerate. |
| Ignoring "etc." / "and so on" |
Explicit signal the list is incomplete. MUST find more. |
| Rename-as-generalize |
Abstraction only produces variations of the example. Re-climb. |
| Over-abstracting on research tasks |
User wants the list. Name concept briefly, then enumerate. |
| Pattern-matching syntax not signal |
"User used commas" is format-matching, not interpretation. |
| Folding competing scopes |
One says "X, Y etc.", another says "scope to X." Enumerate primary requester's concept; flag the rest. |
| "The user specifically said X" |
They said "for example X" — the marker IS the instruction. |
| "I don't want to over-engineer" |
Under-scoping creates rework. Do it once. |
| "Better to be precise than wrong" |
Precision on the example while missing the pattern is precisely wrong. |
| "I'll start with X, they can ask for more" |
Delegation upward. The user expects you to do the thinking. |
| "I'll start broad and narrow down" |
You'll narrow to the examples (concept drift). |
Validation Checklist
After generalizing, verify:
- Concept or rename? Can you generate instances the example didn't give?
- All cases same concept? No mixed categories.
- Reader can't identify user's items. If they can pick them out, restructure.
- Concept stable through enumeration. Every item traces to primary requester's concept.
Example Dialogues
User: "Add input validation to the form fields, for example the email field."
❌ Adds validation only to email.
✅ "I'll add validation to all form fields — email (format), password (strength), username (length/characters), phone (format). Both client and server. Sound right?"
User: "We should handle API errors better, e.g., the 401 on the login page."
❌ Only handles 401 on login page.
✅ "Structured error handling across the app. Login: 401, 429, 500, network failures. Same pattern on signup, password reset, settings — they share the API error surface."
1---2name: generalize3description: Use when a user provides examples prefixed by "for example", "such as", "e.g.", "like", or ends a list with "etc." / "and so on". For architecture/design tasks, abstract first to identify the concept, then enumerate. For research/clarification tasks, enumerate first — the concept organizes the output but the list is the deliverable.4license: CC-BY-SA-4.05---67# Generalizing From Examples89## Overview1011Users describe by listing instances. "Like X" or "for example Y" means they trust you to find the larger concept.1213**Two imperatives; which leads depends on context:**1415| Context | Lead With | Why |16|---|---|---|17| **Architecture / design** | **Abstract first** | Concept determines what to build. "Add monitoring, for example track response times" → Observability, not latency tracking. |18| **Research / clarification** | **Enumerate first** | The list IS the deliverable. Concept organizes it. |1920Enumeration is always produced. The context determines which imperative drives.2122## When to Use2324**Triggers:**2526- Example markers: "for example", "e.g.", "like", "such as", "including", "things like", "something like"27- Softening qualifiers: "or something", "and so on", "etc.", "and things like that"28- Tool-as-category: "use something like Redis" → caching, not Redis29- One concrete case where the domain obviously contains many parallel cases3031**Do NOT use when:**3233- Closure language: "exactly", "specifically", "only", "precisely"34- Closed-set signals: numbered lists, cardinal words ("the three things"), final "and" without "etc."35- Inherently singular scope (one bug, one file, one function)36- You're guessing categories the user wouldn't recognize as related3738## Core Pattern3940### The Inference Ladder41421. **Detect marker** — "for example" = illustrative, not exhaustive432. **Isolate example** — what concept is this an instance of?443. **Identify concept** — what ONE thing do all potential instances share?454. **Infer scope** — architecture: all items across system; research: all items in category465. **Enumerate cases** — always produced, never skipped476. **Confirm boundary** — "Does this cover what you meant?"4849**The ladder is recursive.** Re-climb on your own output. The most common failure is **rename-as-generalize**: your "concept" is just the example in disguise. Test: can you name instances the example didn't give? If not, re-climb.5051### Closure vs. Openness Signals5253The user signals either a **closed set** (complete) or an **open set** (samples). Look for the signal, not the format.5455**Closure (exhaustive — do NOT add):**5657| Signal | Examples |58|---|---|59| Numbered lists / cardinal words | "1. auth, 2. rate limit", "the three things are" |60| Definite quantifiers / finality | "the complete list", "nothing else", "that's it" |61| Closing conjunctions | "X, Y, and Z" (final "and" with NO "etc.") |62| Closure words | "specifically", "only", "exactly", "just" |6364**Openness (illustrative — generalize freely):**6566| Signal | Examples |67|---|---|68| Example markers | "for example", "e.g.", "like", "such as" |69| Softening qualifiers | "or something", "and so on", "etc." |70| Indefinite / hedging | "some", "a few", "maybe", "could include" |71| Tool-as-category | "use Redis" when context is caching strategy |7273**When both present:** primary requester's directive wins. Another stakeholder's dimension is a competing scope definition, not another example — document the conflict, don't resolve by inclusion. Legal requirements (GDPR, SOC2) get flagged separately. **Drift test:** remove the primary requester's message — would your output still make sense? If yes, you've drifted.7475### Abstraction Mechanics7677The concept is the **invariant** across all instances — the underlying capability, not a category label.7879**Abstraction depth:**8081| Example (shadow) | Widening (still shadows) | Concept (the form) |82|---|---|---|83| "track response times" | "monitor all endpoints" | **Observability**: golden signals across all services |84| "use something like Redis" | "pick a cache" | **Caching layer strategy**: session store, query cache, rate limiter |85| "validate email format" | "validate other fields" | **Input integrity**: format, type, range, sanitization, business rules |8687**Acid test:** present the concept without the example — would the user recognize it?8889**Confidence:** 1 example = confirm before enumerating widely. 2–3 = enumerate as educated inference. 4+ = the abstraction IS the request.9091### Enumeration Rules92931. Name the abstract category942. List candidates with user's examples nested within953. Frame additions as inference: "Based on [category], also consider [cases]."964. If abstraction might overreach: "Do you mean [deeper concept]? If so, I'd cover [cases]."975. Let the user trim; don't self-censor. Never replace the user's examples — build outward.9899**Structure by category, not by example.** Example-anchored output means you've appended, not generalized.100101❌ `Caching Research → Priority 1: Redis (user mentioned), Priority 2: Memcached`102✅ `Caching Layer Strategy → Session store: Redis, Memcached; Query cache: Redis (read-replica); Object cache: Hazelcast`103104Test: can a reader identify which items the user mentioned? If yes, restructure.105106## Stop and Re-climb107108| What You're Doing | Why It's Wrong |109|---|---|110| Treating "for example X" as "only X" | The marker signals non-exhaustiveness. X is a starting point. |111| Asking the user to enumerate | The marker means the user expects YOU to enumerate. |112| Ignoring "etc." / "and so on" | Explicit signal the list is incomplete. MUST find more. |113| Rename-as-generalize | Abstraction only produces variations of the example. Re-climb. |114| Over-abstracting on research tasks | User wants the list. Name concept briefly, then enumerate. |115| Pattern-matching syntax not signal | "User used commas" is format-matching, not interpretation. |116| Folding competing scopes | One says "X, Y etc.", another says "scope to X." Enumerate primary requester's concept; flag the rest. |117| "The user specifically said X" | They said "for example X" — the marker IS the instruction. |118| "I don't want to over-engineer" | Under-scoping creates rework. Do it once. |119| "Better to be precise than wrong" | Precision on the example while missing the pattern is precisely wrong. |120| "I'll start with X, they can ask for more" | Delegation upward. The user expects you to do the thinking. |121| "I'll start broad and narrow down" | You'll narrow to the examples (concept drift). |122123## Validation Checklist124125After generalizing, verify:1261271. **Concept or rename?** Can you generate instances the example didn't give?1282. **All cases same concept?** No mixed categories.1293. **Reader can't identify user's items.** If they can pick them out, restructure.1304. **Concept stable through enumeration.** Every item traces to primary requester's concept.131132## Example Dialogues133134**User:** "Add input validation to the form fields, for example the email field."135136❌ *Adds validation only to email.*137✅ "I'll add validation to all form fields — email (format), password (strength), username (length/characters), phone (format). Both client and server. Sound right?"138139**User:** "We should handle API errors better, e.g., the 401 on the login page."140141❌ *Only handles 401 on login page.*142✅ "Structured error handling across the app. Login: 401, 429, 500, network failures. Same pattern on signup, password reset, settings — they share the API error surface."
Arcadi4